The Economic Impact of Depression Treatment in India: Evidence from Community-Based Provision of Pharmacotherapy

The Economic Impact of Depression Treatment in India:
 Evidence from Community-Based Provision of Pharmacotherapy

 Manuela Angelucci∗
 Daniel Bennett†

 June 4, 2022

 Abstract
 This study evaluates the impact of depression treatment on economic behavior in
 Karnataka, India. We cross-randomized pharmacotherapy and livelihoods assistance
 among 1000 depressed adults and evaluated impacts on depression severity, socioeco-
 nomic outcomes, and several potential pathways over 26 months. Pharmacotherapy
 reduces depression severity, especially when paired with a light-touch livelihoods inter-
 vention, with benefits that persist after treatment concludes. The treatments increase
 child human capital investment, particularly for older children, and decrease risk tol-
 erance and the incidence of negative shocks. These findings suggest two pathways
 through which treating depression may reduce the intergenerational transmission of
 poverty.

 JEL: I15, I18

 Keywords: Depression, Health, Poverty

 We received helpful feedback from Vittorio Bassi, Sonia Bhalotra, Leandro Carvalho, Johannes Haushofer, Sylvan Her-
skowitz, Anil Kumar, Emily Nix, Alreena Pinto, Shoba Raja, Gautam Rao, Frank Schilbach, and Scott Templeton. This
research was supported by the Swiss Agency for Development and Cooperation (SDC) and the Swiss National Science Foun-
dation through the Swiss Programme for Research on Global Issues for Development (r4d programme) through the grant
“Inclusive social protection for chronic health problems” (Grant Number: 400640-160374). We also received support from
the Jameel Poverty Action Lab Urban Services Initiative and the University of Michigan. This study is registered as Trial
AEACTR-0001067 in the AEA RCT Registry.
 ∗
 Department of Economics, University of Texas at Austin, mangeluc@utexas.edu.
 †
 bennettd@usc.edu, Center for Economic and Social Research and Department of Economics, University of Southern Cali-
fornia

1 Introduction

Depression is a pervasive and costly illness with a lifetime prevalence of 15-20 percent
(Moussavi et al. 2007, Ferrari et al. 2013, Hasin et al. 2018). It is the fourth largest contrib-
utor to the global burden of disease and the third largest source of years lost to disability
(James et al. 2018). Depression symptoms include anhedonia (the inability to feel pleasure),
pessimism, and disrupted sleep and nutrition. These symptoms may lower productivity (Beck
et al. 2011), reduce the willingness or ability to invest in child human capital (Cummings and
Davies 1994), and affect participation in household decisions (Baranov et al. 2020), thereby
impacting socioeconomic outcomes throughout the household. By addressing these symp-
toms, depression treatment may have health benefits and improve socioeconomic outcomes.
 For developing countries, it is particularly important to understand the economic impact
of depression and find effective and scalable treatments: depression is more prevalent among
the poor and may contribute to poverty and poverty traps (Ridley et al. 2020, Kessler and
Bromet 2013, Haushofer and Fehr 2014). Despite a high need for treatment, the supply of
mental health care is constrained by a shortage of providers in low-income countries (Saxena
et al. 2007).
 Pharmacotherapy may be a useful tool to treat depression in developing countries.
Clinical studies demonstrate the effectiveness of this approach in industrialized countries
(Gartlehner et al. 2017). However, we lack evidence of the feasibility and effectiveness of
community-based provision of pharmacotherapy in developing countries, as well as evidence
of the long-term effects of pharmacotherapy in general. We also lack knowledge of how
mental health care may affect outcomes such as time use, earnings, and investment, and the
pathways through which these effects may occur.
 This paper studies the effects of pharmacotherapy on depression, socioeconomic out-
comes, and possible pathways that may link mental health and economic behavior. We
implemented a community-based cluster cross-randomized trial offering Psychiatric Care
(PC) and Livelihoods Assistance (LA) to 1000 adults (86 percent of whom are female) with
symptoms of mild or moderate depression in a peri-urban region near Bangalore, India. PC
and LA are two commonly available services to treat depression both in this setting and else-

 1

where. Some mental health care providers believe that livelihoods assistance may increase
the effectiveness of pharmacotherapy.
 The PC intervention provided eight months of personalized pharmacotherapy with the
diagnosis and oversight of a psychiatrist from a local research hospital. The LA intervention
consisted of two group meetings to address work-related challenges, followed by personalized
support to help participants find employment or other income-generating opportunities. We
delivered both interventions using the existing local infrastructure: we partnered with a
local NGO that offers these programs to people with mental illness. We measured impacts
on the mental health, time use, and earnings of participants, human capital investment in
children, and consumption, durable goods ownership, and hygiene/sanitation of households,
as well as several potential pathways that could link depression to these outcomes. Forty-four
percent of participants complied with the PC intervention and 68 percent complied with LA.
This level of participation suggests that it is possible to surmount barriers to mental health
treatment such as a lack of awareness and high stigma. We assessed impacts while the PC
intervention was ongoing (our “during” period) and 16-26 months after it began (our “after”
period). The follow-up data allow us to measure the longer-term effects of pharmacotherapy
on mental health and other outcomes, which are largely unknown.
 Both pharmacotherapy arms reduce depression severity. While effect sizes are compara-
ble to the literature, the effect of PC/LA is larger and more persistent. The effect of PC on
depression severity is -0.14 SD (-0.28/-0.001) while PC is ongoing and -0.04 SD (-0.17/0.09)
after PC concludes. The effect of PC/LA is -0.26 SD (-0.39/-0.12) during the PC interven-
tion and -0.24 SD (-0.38/-0.10) afterward.1,2 PC/LA is also more cost effective because the
cost of adding LA to PC is relatively small.
 Neither PC nor LA increases work time or earnings. PC reduces work time by 5.4
hours per week (2.6/8.2) during the PC intervention but this effect dissipates afterward.
By contrast, PC/LA does not reduce work time during the PC intervention.3 Household
 1
 We report average intent to treat effects and provide lower and upper bounds based on 90 percent
confidence intervals in parentheses throughout the paper.
 2
 The impact of LA on depression severity is -0.08 SD (-0.22/0.06) during the intervention and 0.01 SD
(-0.13/0.14) after. The short term impact translates into a decline of 7 percentage points (0.03/13) in the
probability of moderate or severe depression.
 3
 Despite the intention behind the program, LA alone does not have labor market impacts. The lack of
an impact suggests that the LA intervention alone is not sufficient to overcome the barriers to market work

 2

consumption follows a similar pattern: PC significantly reduces consumption during the
intervention but PC/LA does not. The effects of PC and PC/LA on work time and con-
sumption are statistically different during the PC intervention but not afterward. Therefore,
bundling LA with PC has the additional benefit of protecting against some temporary nega-
tive effects of PC. None of the interventions has a statistically significant effect on earnings,
hygiene/sanitation, or durable goods ownership.
 We find large benefits of depression treatment on child human capital investment after
the PC intervention. PC increases investment by 0.18 SD (0.02/0.34) and PC/LA increases
investment by 0.12 SD (-0.09/0.33). Effects are larger and more significant for older children.
Among children who are 12 or older (the age of transition to secondary school), PC increases
investment by 0.44 SD (0.17/0.70) and PC/LA increases investment by 0.40 SD (0.01/0.79).
These results complement Baranov et al.’s (2020) findings that offering psychotherapy to
low-income Pakistani women with perinatal depression increases subsequent child human
capital investment. Effect sizes are comparable to the impact of conditional cash transfers
(Baird et al. 2014), as well as other initiatives to increase student enrollment and attendance
(Evans and Yuan 2020).4
 We consider possible pathways through which depression treatment may affect behavior
and find evidence consistent with a preference pathway, as the treatments increase risk
intolerance and reduce the incidence of negative shocks. Finding that depression treatment
changes preferences is consistent with Bhat et al. (2022), although the specific impact on
preferences varies. Conversely, we can rule out improved cognition and participation in
household decisions as pathways through which depression treatment fosters socioeconomic
improvements in our sample.
 This paper advances several areas of research. We contribute to the studies of the
effectiveness of pharmacotherapy in three ways. First, we establish that a community-based
pharmacotherapy intervention in a developing country is feasible and effective at reducing
symptoms of depression. Therefore, pharmacotherapy may be an additional tool to address
the unmet mental health care needs of the global poor. Secondly, we study the medium-term
among study participants.
 4
 LA also increases child human capital investment. This effect could arise through a small effect of LA
on mental health or through an effect of the interventions on human capital investment via other channels.

 3

effects of a single course of pharmacotherapy, while most studies look at its contemporaneous
effects only. Thirdly, we show that adding LA to PC enhances the effects on mental health
and protects against temporary negative impacts, suggesting that pairing pharmacotherapy
with additional light-touch programs may be cost-effective (Wiles et al. 2016).
 We contribute to the literature on child development, which correlates parental depres-
sion with impaired child development (Cummings and Davies 1994) and lower human capital
investment (Claessens et al. 2015, Dahlen 2016, Shen et al. 2016). While most papers in this
area are descriptive, our study points toward a causal effect of depression on these outcomes.
This finding also contributes to the literature that studies effective interventions to pro-
mote children’s education by identifying an additional demand-side barrier to human capital
accumulation.
 Lastly, we contribute to the literature on the psychology of poverty by exploring the
link between mental health and poverty (Mani et al. 2013, Mullainathan and Shafir 2013,
Haushofer and Shapiro 2016). We do not find evidence that depression treatment increases
productivity, work time, or earnings. While this result is initially surprising, it aligns with
findings by Baranov et al. (2020) and Bhat et al. (2022), who also focus on women in South
Asia. However, we find that depression may reduce investment in child human capital, in-
crease risk tolerance, and thereby expose households to additional negative economic shocks.
These behaviors reduce the future consumption of children and interfere with wealth accu-
mulation (Lybbert et al. 2004, Carter and Barrett 2006), suggesting an intergenerational
link between depression and poverty. A related literature establishes that childhood poverty
leads to adult mental illness (e.g. Persson and Rossin-Slater 2018, Adhvaryu et al. 2019).
Our findings highlight the bi-directional causal relationship between depression and poverty
(Ridley et al. 2020) by showing that adult depression may contribute to future poverty in
children.

2 Setting and Interventions

We conducted this study in a peri-urban region northwest of Bangalore, Karnataka. Our
study area comprises 506 villages and wards (urban jurisdictions) with at least 40 households

 4

within the catchment area of our partner NGO in the Doddaballapur, Korategere, and
Gauribidanur districts. To measure the prevalence and correlates of depression in this area,
we concurrently surveyed a representative sample of adults in an adjacent non-study district.
In this setting, 24 percent of adults aged 18 to 50 have some depression symptoms and
10 percent have symptoms of at least moderate depression.5 Symptoms are more severe
for women, older people, and people with low socioeconomic status, as studies document
elsewhere (Gilman et al. 2002). We elaborate on these patterns in Appendix A.1.
 We study the effects of community-based provision of pharmacotherapy among adults
who screen positive for depression. We collaborated with Grameena Abudaya Seva Samsthe
(GASS), a local social service organization that has worked with people with physical and
mental disabilities since 2001. GASS aims to improve mental health and patient wellbeing
by facilitating psychiatric care and providing livelihoods assistance. To support psychiatric
care, GASS organizes walk-in clinics, sets up appointments, and helps transport people to
health centers. It provides livelihoods assistance by counseling patients about employment
and other earnings opportunities and by helping patients obtain training and small loans as
appropriate.
 The PC intervention provided eight months of free psychiatric care through the Shridevi
Institute of Medical Sciences and Research Hospital. Shridevi is a local private hospital
that offers pro bono care to some patients and sometimes receives patients from GASS.
The initial visit included a diagnosis, an explanation of the significance of mental illness,
and an individualized course of medical treatment. Patients returned for monthly follow-up
visits. The most commonly prescribed anti-depressants were Selective Serotonin Reuptake
Inhibitors (SSRIs). These drugs are generally not under patent and are available inexpen-
sively in India. They are widely used and have relatively few well-tolerated side effects
(Ferguson 2001, Cascade et al. 2009).6 Appendix A.2 discusses ethical considerations. In
 5
 The prevalence of depression in our sample exceeds Sagar et al.’s (2020) estimate of the nationwide
prevalence of 3-4 percent. In part, this pattern may reflect higher depression prevalence in Karnataka than
elsewhere. Moreover, national estimates include a representative share of young people, who have lower rates
of depression than adults.
 6
 Unlike in a clinical trial, participants were aware of their participation in the PC intervention. A portion
of the treatment effect on mental health may arise through a placebo effect. This non-blinded arrangement
realistically characterizes depression treatment in practice. A meta-analysis by Arroll et al. (2005) shows that
treatment with SSRIs is more effective than a placebo in primary care, where the characteristics of patients

 5

addition to treating depression, the PC intervention may raise awareness and salience of
depression in the household, which could lead to additional effects.
 The LA intervention provided two group meetings and personalized livelihoods assis-
tance. The meetings, which lasted three hours each, discussed ways to earn income and deal
with on-the-job challenges. Each meeting had about 30 participants. In the first meeting,
participants had group discussions of their experiences working and earning income, as well
as the challenges they perceived in the labor market. In the second group meeting, facilita-
tors sought to identify suitable livelihoods activities for participants. In subsequent weeks,
staff provided one-on-one assistance to help participants pursue income-generating activities
through job placements, small loans, or training, according to participants’ individual needs
and circumstances. This intervention took place during the first two months of the study.
Although the program was intended to facilitate economic opportunities, the group meetings
may have fostered informal support by bringing participants together (Pfeiffer et al. 2011).

3 Design, Sampling, and Recruitment

The study design and analysis follow the analysis plan that we pre-specified and registered
before collecting follow-up data. Table A1 itemizes and explains our minor deviations from
the analysis plan. We used a cluster-randomized design to cross-randomize psychiatric care
(PC) and livelihoods assistance (LA) by village or ward (urban jurisdiction).7 Figure 1
provides a CONSORT chart for this study. Before starting the recruitment, we stratified the
randomization by district and terciles of a village socioeconomic index based on the 2011
Census of India, for a total of nine strata.8 We then selected 1-2 participants per village. This
and the manifestations of depression often differ from inpatient psychiatric settings. A meta-analysis by
De Maat et al. (2006) shows that pharmacotherapy and psychotherapy are similarly effective on average, and
that pharmacotherapy is effective for treatment of both mild and moderate depression. Around 20 percent
of patients who abruptly discontinue SSRIs experience antidepressant discontinuation syndrome. Symptoms
such as dizziness, fatigue, nausea, and irritability may last for 1-2 weeks (Fava et al. 2015, Gabriel and
Sharma 2017), although evidence regarding this phenomenon continues to evolve (Davies and Read 2019).
Discontinuation symptoms are milder and occur less frequently for patients who receive shorter courses of
treatment (Warner et al. 2006, Eveleigh et al. 2018). GASS organized all visits, transported participants to
their appointments, and monitored patient welfare via home visits throughout the intervention.
 7
 Hereafter we refer to villages and wards as “villages.”
 8
 Socioeconomic index components include village averages of house quality, electrification, latrine use,
and durable good ownership.

 6

design minimized spillovers and cross-arm contamination. Treating few people per village
limited information leakages, protecting patient confidentiality.
 Our partner NGO had limited capacity for both the PC and LA interventions. To
increase statistical power given this constraint, we allocated twice as many participants to
the control arm as to each of the other intervention arms. We ultimately enrolled 395
participants (from 204 villages) in the control arm, 207 participants (from 99 villages) in
the PC arm, 205 participants (from 102 villages) in the LA arm, and 195 participants (from
101 villages) in the PC/LA arm. With these sample sizes, the minimum detectable effect
(MDE) for the comparison of any of the interventions with the control group (e.g. PC/LA
vs. Control) is 0.16 SD in either the “during” or “after” periods. This calculation is based
on the assumptions of 80 percent power and 5 percent significance. For a comparison of two
interventions (e.g. PC/LA vs. PC), the MDE is 0.19 SD. For a test of the complementarity
between the interventions (whether PC/LA = PC + LA), the MDE is 0.28 SD.9 Appendix
A.3 discusses these calculations further.
 We began recruitment in December 2016. We sampled participants through a door-skip
pattern in which the skips were proportional to village size. Once at the household, surveyors
randomly chose an available adult to screen for eligibility. We screened people for depression
symptoms with the PHQ-9 depression severity scale (Kroenke et al. 2001). This nine-item
scale ranges from 0 to 27 and higher values indicate more severe symptoms. The PHQ-9 is
widely validated to screen for depression and measure the response to treatment in India and
throughout the world (e.g., Patel et al. 2008, Manea et al. 2012, Indu et al. 2018). To obtain
a sample of mildly or moderately depressed people, we recruited subjects with PHQ-9 scores
of 9-20.10 In total, surveyors screened 6446 people in order to enroll a study sample of 1000
participants across 506 villages.
 9
 The difference in sample size across arms and periods is small enough that it has negligible influence on
the MDE.
 10
 We initially used a minimum PHQ-9 threshold of 7 before revising the threshold to 9 based on our success
with recruitment. As a result, 8 percent of participants have baseline PHQ-9 scores of 7 or 8. Following
our IRB protocol, we referred people with PHQ-9 scores of 21 or more (indicating severe depression) for
immediate treatment and did not enroll them in the study. To select the people most likely to benefit
from the livelihoods intervention, we did not recruit people who had disabilities that prevented them from
working, who were currently earning more than Rs. 6000 per month, or whose child care duties required
them to remain at home throughout the day. We also excluded pregnant women due the additional risks of
pharmacotherapy during pregnancy.

 7

We did not stratify by gender during recruitment, and 86 percent of participants are
female. This gender ratio is common in other depression studies (e.g. Patel et al. 2017) and
reflects the higher prevalence of depression among women.

4 Data and Measurement

We surveyed respondents five times over 26 months. Round 1 took place at recruitment,
before the start of the interventions. Round 2 occurred four months after recruitment,
midway through the PC intervention and at the end of the LA intervention, and Round 3
occurred eight months after recruitment, around the end of the PC intervention. Round 4
occurred 16 months after recruitment and Round 5 occurred 26 months after recruitment.
We refer to Rounds 2 and 3 as “during the PC intervention” and Round 4 and 5 as “after
the PC intervention” in our analysis below. Figure A2 illustrates the study timeline.
 We study four categories of outcomes: (1) primary outcomes, including depression sever-
ity, work hours, and earnings for participants; (2) child human capital investment; (3) house-
hold consumption, wealth, and hygiene/sanitation; (4) and potential pathways that link de-
pression to the other outcomes. We winsorize monetary values at 5 percent and convert to
2017 values using the Indian consumer price index.
 We measure depression severity using the PHQ-9 scale. The PHQ-9 is not a diagnostic
tool. However, scores of 5-9 roughly correspond to mild depression and scores of 10-20
roughly correspond to moderate or moderately-severe depression, with 88 percent sensitivity
and specificity (Kroenke et al. 2001). Patients with PHQ-9 scores of 10 or more are likely to
have major depressive disorder and are generally referred for medical treatment. We examine
impacts on standardized PHQ-9 scores.
 We measure work time – the time spent on productive activities – from a 24-hour time
diary, which we convert into a weekly value. Productive activities include primary and
secondary jobs, agricultural work, as well as child care, cooking, cleaning, doing laundry,
and fetching water.11 We measure weekly earnings from primary and secondary jobs.
 11
 In addition, we elicit the time devoted to primary and secondary jobs and domestic work in the past
seven days. Estimates using this definition of work time yield similar results. We prefer the time diary
approach because it includes time spent on productive tasks that the respondent may not define as work.

 8

We measure child human capital investment for all children within the household aged
5-18. Outcomes include current school enrollment, days of attendance, hours of homework,
and whether the child currently works for pay. We do not observe any of these variables in
Round 5. We use child-level data for these estimates but we weight by the inverse number
of children per household so that estimates are comparable to other results in the paper.12
 Per-capita consumption is the sum of household food consumption in the past week
(across 23 food groups that are common locally) and expenditures on 13 non-durable non-
food commodities (converted into weekly values from 1 or 2 month recalls) divided by house-
hold size.13 We measure durable goods ownership according to indicators for household own-
ership of nine goods.14 We measure hygiene and sanitation by observing whether there is
open defecation or visible garbage at the respondent’s home, whether the cooking area is
clean, and whether the respondent has visibly dirty hands and fingernails.
 To identify several potential pathways for the socioeconomic impacts of depression treat-
ment, we measure cognitive performance, risk intolerance, subjective wellbeing, and partic-
ipation in household decisions. We assess cognitive performance through three incentivized
tests: Raven’s Progressive Matrices, which estimates fluid intelligence, and forward and
backward digit spans, which measure verbal short term and working memory. We elicit risk
intolerance through items from the Blais and Weber (2006) DOSPERT scale, a generalized
risk self-assessment (Dohmen et al. 2011), and the Eckel and Grossman (2008) incentivized
lottery game.15 We use the five-item Satisfaction with Life Scale to measure subjective well-
being (Kobau et al. 2010). As a measure of participation in household decisions, participants
indicate whether they make household financial and employment decisions alone, with other
 12
 Estimates based on household averages yield similar results. 54 percent of study participants live with
school-aged children and treatment effects on depression are similar regardless of whether school-aged children
are present.
 13
 We include foods that were purchased, produced at home, or received from others. To compute the
value of non-purchased food, we multiply the quantity consumed by median unit values.
 14
 These goods are a chair, a bed, a table, an electric fan, a television, a refrigerator, a bicycle, a motorcycle
or scooter, and a car.
 15
 We measure these variables in Rounds 1-4 only. For the DOSPERT scale items, participants indicate
their willingness to ride a motorbike without a helmet, leave their children unattended for 30 minutes, lend
money to a neighbor, invest 10 percent of annual income in a new business venture, eat spoiled food, and
delay a child’s health care. The first four items are from the original DOSPERT scale and the last two items
are customized to our setting. The incentivized lottery exercise asks participants to choose from a menu of
binary lotteries with payoffs that differ in variance and expected value.

 9

household members, or not at all.
 Since each family of outcomes has multiple variables, we create family-specific indices
by computing the first principal component of the outcomes within each family. This ap-
proach accounts for multiple inference within families. We define the sign of the components
within each group so that larger values have a common interpretation. We also standardize
these indices to ease interpretation. As exceptions to this approach, total consumption is
defined as the sum of food and non-food consumption. For participation in household de-
cisions, we count the number of decisions (across financial and employment decisions) that
the respondent participates in.

5 Treatment Compliance

Across the three arms that received either PC or LA, 65 percent of participants had at
least one psychiatric meeting (for PC) or livelihoods-related interaction (for LA) within
the interventions. Similar proportions of PC and PC/LA participants (45 and 43 percent)
attended at least one psychiatric visit (p = 0.51 for this comparison) according to psychiatrist
records. Similar proportions of LA and PC/LA participants (66 and 71 percent) attended
at least one livelihoods assistance meeting (p = 0.36 for this comparison). Within PC/LA,
31 percent of participants took up both interventions. Figures A3 and A4 further illustrate
intervention compliance.
 91 percent of people who met with a psychiatrist were diagnosed with depression. Pa-
tients who were diagnosed with depression received SSRIs for a median of four months. When
asked in Round 4 to recall drug usage during the PC intervention, 91 percent of participants
report that they took medications either “every day” or “every other day” and 13 percent of
patients continued to take SSRIs after the PC intervention ended. Medication adherence is 8
percentage points higher in the PC/LA arm (p = 0.07). This difference suggests that the LA
treatment may have enabled participants to plan or follow through.16 Among LA compliers,
81 percent attended at least one livelihoods workshop and 47 percent received personalized
 16
 Some patients were also diagnosed with anxiety, pain and high blood pressure, which are common
depression comorbidities (Hirschfeld 2001, Bair et al. 2003, Meng et al. 2012). These patients received
appropriate treatment for these comorbidities (e.g., pain relievers or beta blockers).

 10

livelihoods assistance.17 Appendix A.4 considers the correlates of intervention compliance.
PC and PC/LA compliers are more likely to be men than non-compliers, while LA compliers
are more likely to have better mental health than non-compliers. However, these differences
are not large and compliers and non-compliers do not differ along most dimensions, including
SES and household economic circumstances. Moreover, aside from better mental health in
LA, complier characteristics do not differ across arms. Because the compliance rate and the
characteristics of compliers are similar in PC and PC/LA, differential impacts of PC/LA
relative to PC are unlikely to arise because of differences in intervention participation.

6 Identification and Estimation

We estimate the parameters of the following equation for respondent i in village j and in
round t:

 Yijt = β1 [P Cj · Dt ] + β2 [LAj · Dt ] + β3 [P C/LAj · Dt ] +

 β4 [P Cj · At ] + β5 [LAj · At ] + β6 [P C/LAj · At ] + (1)

 Xij0 β7 + εijt

The variables P C, LA, and P C/LA are indicators for the arms that receive PC only, LA
only, or both PC and LA. D (“during”) and A (“after”) are indicators for Rounds 2 and
3 (while PC was ongoing or had just concluded) and Rounds 4 and 5 (up to 26 months
after the start of the PC intervention). X is a vector of predetermined covariates. The
parameters β1 to β6 identify the Average Intent to Treat (AIT) effects of each intervention
arm under the assumptions that potential outcomes of each treated person are unaffected by
the treatment status of other people and treatment assignment is independent of potential
outcomes. Assigning treatment by village minimizes instances of violations of the first as-
sumption through spillovers such as social interactions, while treating 1-2 people per village
minimizes village-level general equilibrium effects. Random assignment should ensure that
the second assumption holds.
 17
 Nobody in the control group sought treatment through GASS. It is possible but unlikely that control
participants sought treatment elsewhere; most people with mental disorders go untreated in this setting.

 11

We test whether PC and PC/LA have the same effects (β1 = β3 and β4 = β6 ) and
whether there are no complementarities between P C and LA (β3 − β1 − β2 = 0 and β6 − β4 −
β5 = 0). Moreover, we test that the treatment effects do not differ by arm (β1 = β2 = β3 and
β4 = β5 = β6 ) and that the other pairwise effects are identical (e.g., β1 = β2 , and β3 = β2 ).
We use OLS and cluster standard errors by village.
 We estimate ANCOVA and LASSO versions of this specification for all outcomes. Under
ANCOVA, X includes the baseline dependent variable and strata and time dummies.18 The
LASSO approach uses the post-double-selection method of Belloni et al. (2014) to choose
covariates. When these approaches yield similar estimates (the majority of cases), the text
describes the ANCOVA estimates. Otherwise, we note the discrepancy between the two
estimates.19
 Table 1 shows baseline summary statistics of key outcome variables and covariates by in-
tervention arm. Columns 2-3 show the control mean and standard deviation of each variable.
Columns 4-9 show the mean difference between each intervention arm and the control arm,
along with p-values (based on village-clustered standard errors) that indicate the statistical
significance of these differences. Finally, Column 10 provides the p-value for the joint test of
significance of the three intervention arms relative to control. Most outcomes are balanced
across intervention arms in Round 1, and we cannot reject that the variables in the table are
 18
 Our analysis plan prescribes using an ANCOVA specification for outcomes with low serial correlation
and a difference-in-difference specification for outcomes with high serial correlation (McKenzie 2012). In
practice, all outcomes have serial correlations below 0.3, except for the durable goods index, which has serial
correlation of 0.53. Therefore, we use ANCOVA to streamline the analysis. Difference-in-difference estimates
closely resemble ANCOVA estimates and are available from the authors.
 19
 For the lasso regression, we allow the estimator to select from the following list of baseline covariates:
strata indicators, round indicators, gender, marital status, education, scheduled caste/tribe, literacy, house-
hold size, PHQ-9 score and components, PHQ-9 < 10 indicator, PHQ-9 < 5 indicator, GAD-7 (anxiety)
score and components, activities of daily living index and components, time use (all work, paid work, unpaid
work, sleep, leisure, and job search hours), per capita household non-durable consumption and expenditures
(total, food, non-food, clothes for children, medical), sanitation/hygiene index and components, older child
human capital index and components, young child health index and components, per capita net savings and
components, durable goods index and components, risk intolerance index and components, negative shock
index and components, cognition index and components, subjective wellbeing index and components, par-
ticipation in household decision and components. This list includes the baseline values of all outcomes in our
analysis. Child human capital regressions also include child-level covariates: an indicator that the individual
is the child of the study participant, the baseline human capital index and components, and age and gender
dummies. To avoid dropping observations, we include indicators for missing values of all covariates and then
set missing values to zero. The algorithm chooses the baseline dependent variable or some of its components
in 76 percent of cases. All specifications choose at least some time dummies and 36 percent of specifications
select at least some strata dummies. The algorithm selects a median of nine covariates.

 12

jointly balanced (p = 0.21). However, the table shows that PHQ-9 scores are imbalanced,
which could contribute to follow-up differences in this or other outcomes. To address this con-
cern, we also estimate a version of all regressions that uses entropy weights to impose balance
across arms in the first three moments of the PHQ-9 distribution (Hainmueller 2012, Hain-
mueller and Xu 2013). Estimates are robust to weighting, and weighted estimates (available
from the authors) are generally similar to unweighted estimates.
 The last row of Table 1 shows that, overall, attrition does not vary systematically by
arm. However, when we examine attrition by arm, we find a higher attrition rate in the PC
arm for Round 5. Appendix A.5 considers this issue and concludes that differential attrition
does not affect the results we present below.

7 Impacts on Participants

7.1 Depression Symptoms

Table 2 shows treatment effects on depression symptoms. Both pharmacotherapy arms
improve mental health to some extent. However, the impact of PC/LA is significantly larger
and more durable: PC/LA reduces the PHQ-9 score by 0.26 SD (-0.39/-0.12) during the
PC intervention and by 0.24 SD (-0.38/-0.10) afterward, while PC alone reduces the PHQ-9
score by 0.14 SD (-0.27/-0.001) during the PC intervention and by 0.04 SD (-0.17/0.09)
afterward. These effect sizes are consistent with the literature, as we discuss in Appendix
A.6.
 Tests of coefficient equality confirm that the impact of PC/LA is significantly larger than
the impact of PC or LA alone, and fail to reject that PC and LA have similar effects. The
impact of PC/LA is generally larger than the sum of the impacts of PC and LA, consistent
with a complementarity between these interventions. However, this difference is generally
statistically insignificant (0.10 ≤ p < 0.28). LA has smaller effects than the other arms:
-0.08 SD (-0.22/0.06) during the PC intervention and 0.01 SD (-0.13/0.14) afterward. The
differences between the effects of LA and PC are not statistically significant.
 To quantify the differential impact of PC/LA over PC, we compute the total reduction

 13

in PHQ-9 × months over the study period for each arm.20 Under this metric, PC/LA is 3.5
times more effective than PC. Since PC/LA costs just 5 percent more than PC alone ($232
versus $221 per study participant), bundling PC and LA improves the cost effectiveness in
terms of reducing depression symptoms. Appendix A.7 describes this exercise in more detail.
 Two figures provide more information about the treatment effects on mental health.
Figure 2 plots the PHQ-9 densities by arm during and after the PC intervention. Depression
symptoms decrease throughout the support both during and (to a lesser extent) after the
intervention. As noted, impacts are largest for PC/LA participants. Figure 3 plots average
PHQ-9 scores by arm and round. The gap between treatment and control is largest for the
PC/LA arm in every round and smaller, but still positive, for the other arms. It gradually
decreases over time as mental health improves in the control group. This pattern is consistent
Spijker et al.’s (2002) finding that depression symptoms diminish over 1-2 years for most
people but persist for 10-30 percent of patients.
 Appendix A.8 provides effects on the probability of no moderate or severe depression
(PHQ-9 < 10) and no depression (PHQ-9 < 5). These estimates are helpful for comparative
purposes since they are commonly reported in the literature. Table A2 and Figure A5 show
that the results for these outcomes are qualitatively similar to our main results: both PC
and PC/LA reduce the frequency of mild and moderate depression, but PC/LA has bigger
and more long-lasting impacts. We also find that LA has a modest impact on the frequency
of moderate or severe depression.
 Appendix A.9 estimates heterogeneity in the impact on mental health by baseline gender,
age, socioeconomic status, PHQ-9 score, physical health, cognition, and exposure to negative
shocks during childhood. These estimates appear in Figure A6. Both PC/LA and PC have
larger effects for people with worse physical health. PC/LA is more effective for people with
many childhood shocks. We do not find significant heterogeneity in the impact of LA or in
the impacts of any of the interventions along other dimensions.
 Figure A7 in Appendix A.10 estimates impacts on the GAD-7 anxiety score and an index
of activities of daily living (ADL). The PC/LA intervention significantly reduces anxiety
 20
 We multiply the “during” period estimates in Column 1 of Table 2 by eight months and the “after”
period estimates in Column 1 by eighteen months.

 14

while the other interventions do not have statistically significant effects. The impact on
activities of daily living varies by arm: PC/LA statistically increases the ADL index, PC
decreases it during the intervention, and LA does not have statistically significant effects.

7.2 Work Time and Earnings

Table 3 shows that PC/LA and PC have different treatment effects on weekly work time and
earnings while the PC intervention is ongoing. The effect of PC on work time is -5.4 hours
per week (-8.2/-2.6) and the effect on earnings is -65 rupees per week (-155/24), a 10 percent
decrease in both outcomes. By contrast, the effect of PC/LA on work time is 1.1 hours
per week (-1.7/3.8) and the effect of PC/LA on earnings is 38 rupees per week (-63/139).
The difference between the effects of PC/LA and PC in the “during” period is significant
for work time (p = 0.001) but not for earnings (p = 0.12). Figure A8 shows that there is a
concurrent increase in sleep and leisure time in the PC arm, as we discuss in Appendix A.11.
This pattern suggests that PC may reduce work time by increasing the marginal utility of
leisure or self-care. Alternatively, mental health stigma might reduce either labor supply
or demand (Corrigan et al. 2001, Bharadwaj et al. 2017). The significant negative impact
of PC and the significant difference between PC and PC/LA are not present at follow-up,
at which point the effects are negative for both arms, but closer to zero and statistically
insignificant.21
 In sum, our results suggest that pharmacotherapy does not increase the time spent
on productive activities in our sample. 86 percent of our study participants are female,
and low female labor force participation in India may weaken the labor market impacts in
this setting. Baranov et al. (2015) and Bhat et al. (2022) also find no long-term effects of
psychotherapy on labor market outcomes among all or mostly female samples in South Asia.
By contrast, Patel and Kleinman (2003) and Patel et al. (2017) find that mental health
care reduces self-reported work absenteeism and Lund et al. (2018) find that various mental
health interventions have positive effects on employment.
 The LA intervention has no statistically or economically significant effects on work time
 21
 PC/LA reduces work time under the ANCOVA specification (-6.2/-0.4) but has a statistically insignifi-
cant effect under LASSO (-6.3/0.6).

 15

and earnings. In the “during” period, the effect of LA on work time is -1.0 hours per week
(-3.8/1.8) and the effect on earnings is -33 rupees per week (-135/69). In the “after” period,
the effect of LA on work time is -1.5 hours per week (-4.7/1.7) and the effect on earnings is
48 rupees per week (-55/150). There are likely multiple barriers to increasing work time and
earnings for our sample. Our findings suggest that neither mental health care nor livelihoods
assistance are sufficient to overcome these barriers.

8 Impacts on Children and the Household

Impacts on child human capital investment appear in Table 4. Most effects are not sta-
tistically significant in the “during” period, although PC increases investment by 0.13 SD
(0.01/0.26) under the LASSO specification.22 Effects of PC/LA and PC are significantly
different from each other, consistent with patterns for several other outcomes. In the “af-
ter” period, all coefficients are positive and are not significantly different from one another.
However, only the impact of PC is statistically significant under both specifications. Under
ANCOVA, PC has an effect of 0.18 SD (0.02/0.34), while PC/LA has an effect of 0.12 SD
(-0.09/0.33) and LA has an effect of 0.11 SD (-0.08/0.30).23 We also collected health and
anthropometric data for children under age five. However, only 85 study participants resided
with young children who provided measurements. Estimates for these outcomes (available
from the authors) are imprecise but align with the positive impacts on child human capital
discussed here. Finding that the LA arm, which had minimal mental health improvements,
also increases child human capital suggests that the treatments may affect human capital
investment also through channels other than improved mental health. For example, the in-
terventions may raise awareness of mental health within the household or lead the household
to reconsider important economic choices.
 Next, we examine heterogeneity by the median age of 12, which corresponds to the
 22
 The impact on child human capital investment may occur with a lag because enrollment typically occurs
at the beginning of the academic year. In addition, school attendance and homework time are likely to be
inelastic among non-enrolled students and among all students during periods when school is not in session.
Enrollment, attendance, and homework maybe unresponsive in Round 2 because it occurred during the same
academic year as Round 1. In addition, attendance and homework may be unresponsive in Round 3 because
school was not in session for many students at that time.
 23
 Figure A9 shows impacts on the components of the child human capital investment index.

 16

transition to secondary school. Estimates are small and statistically insignificant for younger
children. For older children, PC/LA has an impact of 0.40 SD (0.01/0.79), PC has an impact
of 0.44 SD (0.17/0.70), and LA has an impact of 0.32 SD (0.01/0.64) in the “after” period,
with slightly larger estimates under LASSO. The effects differ significantly by child age for
PC/LA (p = 0.09) and for PC (p = 0.02), but not for LA (p = 0.18). This pattern may
reflect a ceiling on the potential impact for younger children. For example, for children who
are 12 or younger in the control group, 94 percent are enrolled and 0.5 percent work for
pay across Rounds 1-4. By comparison, 85 percent of children over 12 are enrolled and 11
percent work for pay.
 To benchmark these impacts, we compare our estimates for enrollment with the impacts
of both educational interventions on enrollment from Evans and Yuan (2020) and conditional
and unconditional cash transfers from the meta-analysis by Baird et al. (2014). We find that
our estimates are within the range of both sets of outcomes, suggesting that these effects are
economically relevant.
 Figure 4 shows treatment effect heterogeneity in the “after” period by several additional
characteristics, including the child’s gender, relation to the study participant and baseline
human capital, as well as the study participant’s baseline depression severity and gender.
There is a significantly larger effect of PC/LA for boys and children with high baseline human
capital investment, but other differential effects are statistically insignificant.
 Figure 5 shows that the interventions have no statistically significant impacts on hygiene/
sanitation, durable good ownership, or household consumption. An exception to this pattern
is that PC significantly reduces per capita household consumption in the “during” period.
A concurrent decline in per capita household income for the PC arm may be responsible for
this effect.24 Appendix A.13 shows impacts on the components of these indices.
 24
 PC reduces per capita household income by Rs. 44 (5/83) and reduces per capita household consumption
by Rs. 59 (20/96) in the “during” period. No other arms have statistically significant effects on this outcome
either during or after the PC intervention.

 17

9 Potential Pathways

This section considers four pathways that may link depression and depression treatment to
socioeconomic outcomes: risk intolerance, subjective wellbeing, cognitive performance, and
participation in household decisions. Figure 6 shows treatment effects on these outcomes.
 The interventions increase risk intolerance, although the timing and significance of the
effects vary across arms. PC increases risk intolerance by 0.25 SD (0.05/0.43) and PC/LA
increases risk intolerance by 0.18 SD (-0.01/0.38) in the “after” period, while LA increases
risk intolerance by 0.16 SD (0.01/0.26) in the “during” period. This pattern is consistent
with a preferences pathway in which treatment reduces anhedonia and pessimism. These
common depression symptoms reduce the actual and expected marginal utility of consump-
tion, which may decrease the perceived return on human capital investment, as well as the
desire to minimize the incidence of negative shocks.25 This pathway is also consistent with
the relationship between chronic illness and the marginal utility of consumption (Finkelstein
et al. 2013), as well as the negative association between poverty and the enjoyment of various
activities (Schofield and Venkataramani 2021).
 Consistent with this interpretation, we find that the pharmacotherapy interventions
reduce the incidence of negative shocks. Figure A16 shows that PC/LA reduces the incidence
of negative shocks by 0.14 SD (0.01/0.28) and PC reduces the incidence of negative shocks
by 0.11 SD (0.02/0.24) in the “after” period.26 These findings suggest that, in addition to
suppressing human capital investment, depression may also perpetuate poverty by exposing
people to additional shocks that prevent wealth accumulation (Lybbert et al. 2004, Carter
and Barrett 2006).
 Next, we consider the impacts on subjective wellbeing, which is a proxy for utility. Since
depression is emotionally painful, we may expect depression treatment to improve wellbeing
 25
 This logic is commonly applied to mortality risk in order to elicit the value of statistical life (e.g. León
and Miguel 2017). For a reduction in the marginal utility of consumption to have this effect, the cost of
taking action must not decrease commensurately. Fatigue and impaired cognition, which are also depression
symptoms, likely increase the cost of taking action.
 26
 Since “an illness lasting at least one month” is an element of the negative shock index, the interventions
could mechanically improve the index by reducing the incidence of depression. We investigate this possibility
by excluding the illness component and find results that are robust and very similar to the estimates in Figure
A16. Estimates are available upon request.

 18

(Smith et al. 2020). However we find that the interventions reduce subjective wellbeing. The
effect size varies by intervention arm but ranges from -0.18 to 0 SD, despite the observed
improvements in mental health in Table 2. This pattern suggests that the interventions
may change aspirations, expectations, or reference points, which is consistent with findings
by Adhvaryu et al. (2020) that objective improvements in circumstances that fall short of
expectations reduce life satisfaction.
 Depression could also change behavior by affecting cognition. Figure 6 shows an impact
of PC/LA on cognitive performance of -0.16 SD (-0.30/-0.02) and an impact of PC on
cognitive performance of -0.19 SD (-0.32/-0.06) in the “after” period. The lack of a positive
effect rules out that improved cognition is a pathway through which better mental health can
lead to improved socioeconomic outcomes for our sample. Appendix A.14 discusses possible
explanations for this finding.
 Finally, we consider impacts on participation in household decisions. We do not find
evidence for this channel: most estimates are small and statistically insignificant both during
and after the PC intervention. An examination of the components of this index in Figure
A18 suggests a shift toward joint rather than individual decision-making under PC.

10 Joint Significance and Treatment Complementarities

This section tests whether the interventions have effects that are jointly significant across
the eleven main outcomes of our analysis.27 We use the “omnibus” test proposed by Young
(2019) to examine whether the interventions have significant. We reject the null hypothesis
that the three interventions are jointly insignificant (p < 0.001). Implementing this test
separately by intervention arm, we reject the hypothesis of no effect of PC/LA (p = 0.001)
and of PC (p = 0.001) but fail to reject the hypothesis of no effect of LA (p = 0.22).
Therefore we conclude that both pharmacotherapy interventions have significant effects.
 Next, we re-estimate the effects on these eleven outcomes as a system of seemingly
unrelated regressions (SUR) to test the additional hypotheses described in Section 6 jointly
 27
 These outcomes are the PHQ-9 score, weekly work time, weekly earnings, child human capital investment,
hygiene/sanitation, durable goods ownership, per-capita consumption, risk intolerance, subjective wellbeing
cognitive performance, and participation in household decisions.

 19

across outcomes. When comparing the effects of PC and PC/LA (H0 : P C = P C/LA),
we reject equivalence in the “during” period (p = 0.001) but not in the “after” period
(p = 0.37). We reject equivalence if we pool time periods (p = 0.001). When testing for “no
complementarity” in the effects of PC and LA (H0 : P C/LA = P C +LA), we find evidence of
complementarity in the “during” period (p = 0.001) but not in the “after” period (p = 0.74).
We reject this hypothesis if we pool time periods (p = 0.02). These findings reaffirm our
conclusions that pairing LA with PC leads to significantly different outcomes and that LA
appears to temper several transitory negative effects of PC in the “during” period.28

11 Discussion

There is an urgent need for mental health care in India and other developing countries.
In a representative survey we conducted adjacent to the study area, 24 percent of adults
had at least mild depression symptoms and depression was strongly correlated with low
socioeconomic status. Although the Mental Health Care Act of 2017 creates a legally-
binding right to mental health care in India (Duffy and Kelly 2019), only 15 percent of
people with depression in India receive care (Gautham et al. 2020). Evidence regarding the
effectiveness of depression treatment in low-income settings is limited (Patel et al. 2007). The
impact of treatment may differ across developed and developing countries due to disparities
in health care access and quality, the severity of depression, the prevalence of different types
of depression (Harald and Gordon 2012), awareness of mental illness, stigma, and treatment
compliance.
 Psychotherapy and pharmacotherapy are the leading approaches to depression treat-
ment. While studies have shown the utility of psychotherapy as a way to provide depression
care to poor people in developing countries (Baranov et al. 2020, Haushofer et al. 2020, Patel
et al. 2017, Barker et al. 2021), research has not explored the effectiveness of community-
based pharmacotherapy. Since it requires fewer personnel than psychotherapy, pharma-
 28
 The SUR approach allows us to test the remaining Section 6 hypotheses jointly. We fail to reject the
hypothesis that the effects of PC and LA are equal (p = 0.16 overall, p = 0.35 in the “during” period,
p = 0.29 in the “after” period). We reject the hypothesis that the effects of PC/LA and LA are equal overall
(p = 0.07) and in the “during” period (p = 0.03) but not in the “after” period (p = 0.26). We reject the
hypothesis that the three interventions have equal effects overall (p < 0.001) and in the “during” period
(p < 0.001) but not in the “after” period (p = 0.19).

 20

cotherapy may be a valuable tool to treat depression in low and middle income countries,
where mental health specialists are scarce (Saxena et al. 2007).
 In our trial, we find effects on depression symptoms that align with the clinical literature
(Gartlehner et al. 2017). While treating depressed adults increases child human capital in-
vestment, it also has some negative impacts, most of which are transitory. Pairing livelihoods
assistance with pharmacotherapy increases the size and duration of the mental health ben-
efit, preserves the positive effect on child human capital investment, and safeguards people
against several of these negative effects. Adding livelihoods assistance increases intervention
costs by only 5 percent.
 Future research should investigate the complementarity between pharmacotherapy and
livelihoods assistance and whether other inexpensive light-touch interventions enhance the
benefit of mental health care in a similar way. Since LA does not directly increase work time
or earnings, features other than job-related benefits of LA may impact mental health. The
group and individual social interactions that occur under LA may have enabled participants
to receive emotional support from like-minded peers. Higher medication adherence among
the PC/LA participants also suggests that LA may have improved the ability of participants
to plan or follow through. Moreover, LA may have helped participants overcome the stigma
of receiving mental health care by supplying a “reason” for participating without admitting
to mental illness.
 Finding that treating adult depression increases child human capital investment suggests
that the well-known correlation between parental depression and child development is at least
partially causal. The magnitude of this impact is large. Therefore, this finding shows that
adult mental health may be an important demand-side constraint on child human capital
accumulation.
 Finally, our study suggests that depression treatment may change preferences by in-
creasing risk intolerance, consistent with an increase in the marginal utility of consumption.
This finding of an impact on preferences is broadly consistent with the evidence from Bhat
et al. (2022), although they find effects on altruism and patience, rather than risk tolerance.
This divergence between our findings could reflect a difference between psychotherapy and
pharmacotherapy.

 21